Palette, a 20B open model, ranked first on two of three PresentBench domains for slide decks

Palette is a small open-weight model that builds slide decks in a set design style. It was scored on a benchmark called PresentBench, and it came first in two of the three domains measured.

The scores

Palette ranked first on Economics and Education and third on Talk. Its three-domain mean was second: 59.7 against NotebookLM's 60.8. Baseline numbers come from the PresentBench paper.

Imagine you must keep a board deck inside your own building. Would a second-place model you own beat a first-place one you rent?

Why bother with a small model

Large closed models make good decks, but they keep falling back on their own house look, so a batch of decks can feel samey. Some content is also confidential and cannot leave a company's systems.

A model you own, running on your own servers, answers both concerns. The trouble is that small models tend to produce flawed slides: headings that blend into the backdrop, words spilling past the edge, and shapes sitting on top of one another.

Two roles, one model

Palette divides the work as a human team would. A designer role plans the deck and writes a readable brief: palette, fonts, and each slide's layout and content. A coder role then writes the code that draws one slide from that brief.

Both roles run on a single LoRA adapter, a light add-on that teaches a skill without rewriting the base model. Since each slide's context is settled beforehand, the slides can be coded at the same time, making that stage about 6-7 times faster. A person can also edit one slide's layout in the brief without redoing the deck.

Training data that explains itself

No public collection of deck plans and slide code exists, so the team generated one with a large open model. It covers around 1000 decks and roughly 10,000 slides, following the IBM Design Language.

Each example includes its reasoning, including an option that was rejected. One imagined case: a chart beats a table when the point is a trend, not exact figures. Every deck was rendered and checked for overlaps, and broken slides were repaired and their reasoning rewritten.

Inventing facts was the hardest failure. Prompt rules did not stop it. Fixing the reference decks the teacher model saw did.

A checker that cannot make things worse

Users can steer a deck with short directives in double square brackets, such as a request to make a slide a timeline. Twelve directive categories were in the training recipes.

Every finished deck passes through a harness. Simple rules fix known errors, a geometry checker flags overlap and overflow, and a separate editor model rewrites flagged slides. A rewrite is kept only if the checker confirms it is better. Across 131 decks, the harness added 3.5 points on average.

Limits: PresentBench does not test whether a deck follows a specific design language, and the team knows of no benchmark that does. Palette's plans were drafted by a separate frontier model. It trails on narrative Talk decks. The team also wants to close the gap on figure-heavy decks, and bringing your own templates without retraining is still being explored.